Hearing certain natural sounds (e.g., the croak of a frog) appears to automatically activate conceptual knowledge, enabling the perceptual system to quickly identify objects in the surroundings (e.g., the presence of a frog). Learning such cross-modal associations represents a crucial prerequisite for mediating interactions with the environment. In humans, conceptual representations can also be activated via language (e.g., “frog”). However, unlike natural sounds, linguistic symbols are categorical, making them more suited to activate semantic information in a format that transcends within-category differences. It remains unclear whether phylogenetically young systems like language exert effects on perception similar to natural sounds, and what brain dynamics might support such effects.

Conceptual representations activated by auditory cues have been shown to interact with the visual system in different ways. For instance, hearing words and natural sounds can rapidly drive visual attention toward specific entities in a scene (Huettig and Altmann, 2007); facilitate the recognition and discrimination of congruent object categories (Boutonnet and Lupyan, 2015; Edmiston and Lupyan, 2015); lower the detection threshold for ambiguous objects (Lupyan and Ward, 2013; Ostarek and Huettig, 2017); and even cause sensory illusions (Dils and Boroditsky, 2010). While this body of evidence suggests that both linguistic and non-linguistic cues activate content-specific representations, it is less clear whether these cues activate the same representations. Studies directly targeting this issue have often reported a “label-advantage” effect, that is, a facilitation when object recognition is preceded by words compared to non-linguistic cues (Edmiston and Lupyan, 2015; Lupyan and Thompson-Schill, 2012). This effect suggests that language provides a particularly powerful tool to enhance visual processing.

To achieve these facilitatory effects on visual perception, linguistic categories could theoretically follow two possible pathways (Simanova et al., 2016). Language might not bias perceptual processes at early levels but rather interact at later semantic or categorical decision-making stages (Firestone and Scholl, 2014; Gleitman and Papafragou, 2005; Klemfuss et al., 2012). On an alternative account, words could affect visual processing by setting categorical priors that alter early perceptual processing (Boutonnet and Lupyan, 2015; Kok et al., 2014; Thierry et al., 2009). Support for the latter account comes primarily from EEG studies showing that better recognition of images preceded by congruent words was associated with modulations of early event-related potentials (ERPs) such as the P1 (Boutonnet and Lupyan, 2015; Noorman et al., 2018) – putatively considered an electrophysiological index of low-level visual processes (Spehlmann, 1965). Yet, these ERP experiments targeted the perceptual consequences of language cues on visual object recognition i.e., they focused on time interval following the visual stimulus. The mechanisms that could explain prestimulus effects of language on visual perception remain largely unknown¹.

Analysis of oscillatory activity provides an excellent opportunity to study language-driven prestimulus modulations during visual object recognition. Based on previous human and animal studies, candidate mechanisms to carry sensory representations are low-frequency oscillations in the alpha/beta-band (Arnal and Giraud, 2012; Bastos et al., 2012; Bastos et al., 2015; Michalareas et al., 2016). Rhythmic brain activity in these frequency bands has been suggested to play a large variety of roles in top-down processing, which could be crucial to support visual object recognition. For example, alpha synchrony has been associated with filtering of task irrelevant information and enhancement of neural representations during tasks involving attention, prediction, mental imagery and working memory (Hari et al., 1997; Jensen et al., 2002; Mayer et al., 2015; Mo J. et al., 2011). This mechanism could be crucial to speed up object recognition by silencing neural populations encoding irrelevant object categories and activating those linked to the target object. Similarly, beta oscillations have been implicated in perceptual expectations (Arnal and Giraud, 2012), online maintenance of cognitive states (Bressler and Richter, 2015; Engel and Fries, 2010) and (re)activation of task-specific cortical networks (Spitzer and Haegens, 2017). Within the context of visual object recognition, this process can facilitate the activation of network states associated with the target object, thereby speeding up its recognition after the image is presented. Based on these findings, we hypothesized that any object recognition advantage for spoken words over natural sounds would be associated with a difference in cortical alpha/beta dynamics.

In the present study, we used a cue-picture matching task to test the hypothesis that language enhances visual object recognition by setting categorical representations via the modulation of alpha/beta oscillations. In contrast to previous studies, we (i) focused on the time interval preceding the onset of the visual object, targeting top-down signaling directly; and (ii) included words from participants’ first (L1) and second (L2) languages, to assess whether the previously reported label advantage extends to language systems acquired later in development. We hypothesized that, if the label advantage arises because words provide refined categorical representations to the visual system, then any differences in object recognition cued by words vs. natural sounds should be associated with modulations of oscillatory alpha/beta dynamics before the onset of the target picture.

Several studies have reported that spoken words can boost visual recognition of object categories, but the neural mechanisms underlying such facilitation are not well established. It has been suggested that effects of language on visual perception arise at early stages of sensory processing; specifically, via the amplification of category-specific representations in sensory regions. In the present study, we investigated the prestimulus effect of language on visual perception, testing the hypothesis that neural oscillations can serve as mechanisms to carry language-generated representations about incoming object categories.

To test this hypothesis, we used EEG to measure prestimulus brain activity and characterize the oscillatory dynamics underlying the label-advantage in object recognition. We reasoned that, if objects are recognized faster because spoken words provide more refined categorical representations than natural sounds, then these cues should differentially modulate prestimulus oscillatory activity in the alpha and beta bands.

We first replicated the previously reported label-advantage and showed that this behavioral effect persisted even when words were presented in a second language. This suggests that verbal symbols deploy more accurate knowledge representations than natural sounds against which incoming inputs can be compared. Importantly, the reported behavioral advantage for spoken words was associated with an increase in the power of alpha and beta rhythms in the time interval between the offset of the cue and the onset of the target object. Such synchronization points to a possible functional role for alpha and beta neural rhythms in the label advantage in object recognition.

While the widespread distribution of significant clusters – including numerous electrodes across the whole scalp – might suggest that the reported effect reflects global alpha/beta network states, the contrasts between conditions indicate that the effect was particularly pronounced over posterior electrodes. This was particularly evident in the alpha frequency-band, suggesting that alpha waves may reflect local oscillatory states originating in occipital regions, in line with findings from the monkey literature (Mo J. et al., 2011).

Enhancement of alpha oscillations in occipital regions has been largely reported when top-down knowledge is directed by a cue toward a specific feature (Bollimunta et al., 2008; Mo J. et al., 2011) or direction (Snyder and Foxe, 2010; Worden et al., 2000). At least two non-mutually exclusive theoretical accounts have been advanced to explain this effect. Some recent proposals posit that enhancement of neural alpha synchronization in task-relevant regions leads to excitatory effects reflecting selective amplification of neural representations of object categories (Klimesch, 2012; Mo J. et al., 2011; Palva and Palva, 2007; Van Kerkoerle et al., 2014), which could in turn facilitate the recognition of the incoming object. For instance, M/EEG studies have reported that alpha power increases in grapheme-processing regions with the predictability of letter identity (Mayer et al., 2015); and in the posterior cortex when meaningful hints precede the discrimination of ambiguous images (Samaha et al., 2018). Similarly, biophysical models indicate that enhancement of prestimulus alpha waves can improve detection performance by increasing the excitation of pyramidal cells, rendering the network state less stable and thus facilitating the activation of a stimulated assembly (Lundqvist et al., 2013). One possibility is that the reported modulations on alpha activity can serve as a mechanism to carry language-generated representations about the structure of visual objects.

Another prominent view is that enhanced alpha power reflects states of inhibition and filtering of task-irrelevant information (Jensen and Mazaheri, 2010; Klimesch et al., 2007). For instance, when attention is directed toward a target on one side of space, posterior alpha-band power increases at electrodes over the hemisphere ipsilateral to the target (Thut et al., 2006; Worden et al., 2000). According to this view, increased alpha oscillations reflect suppression of cortical areas not involved in the task. The alpha effect in our study was right-lateralized and might reflect the inhibition of right-posterior regions to gate sensory information processing to the left-posterior network, where language-perception interactions usually take place (Mo L. et al., 2011). However, this interpretation would also predict alpha desynchronization over left-posterior regions to increase excitability and enhance stimulus processing (Jones et al., 2000; Klimesch et al., 2007). Since we did not find any evidence for the latter effect, we consider it unlikely that alpha synchronization acted as an inhibitory filter in the current study.

A novel result of our study in contrast with similar earlier studies was the differential beta-band modulations that resulted from spoken word vs. natural sound cues. Recent proposals suggest that beta oscillatory activity reflects endogenously driven transitions from latent to active cortical representations of object categories (Spitzer and Haegens, 2017), as well as the binding of neurocognitive network elements associated with a given neural representation (Bressler and Richter, 2015). Under these accounts, beta synchronization provides “a flexible scaffolding that sets up functional neuronal ensembles through temporary synchronization of content-coding cell populations” (Spitzer and Haegens, 2017). In the context of visual object recognition, language-driven beta waves can reactivate neurocognitive networks associated with the target object, enhancing recognition after the image is presented. We speculate that the difference in beta modulations for spoken words vs. natural sounds may reflect a difference in the content of the (re)activated conceptual states – and more importantly, in the amount of retrieved conceptual dimensions, e.g., the size of the neurocognitive network state (Bressler and Tognoli, 2006). Behavioral and eye-tracking experiments have indeed shown that spoken words activate a rich network of features during lexical processing (Huettig and Altmann, 2007). Consequently, processing words might lead to the retrieval of knowledge dimensions that go beyond the purely sensory features of objects, such as conceptual, grammatical, and lexical information. This is partially in line with human and monkey studies showing that beta synchronization carries supramodal information about object categories (Wutz et al., 2018).

We recognize that the interpretations above regarding the role of alpha and beta oscillations should be considered with caution, particularly in relation to specific frequency effects and the distinct roles of alpha and beta rhythms. In fact, the similar power dynamics observed for alpha and beta in this study suggest that they may serve similar functions and support the same underlying mechanisms. The coupling of these frequencies has been previously reported during naturalistic language comprehension, where their power has been shown to similarly encode high-level linguistic computations, such as dependency-building (Zioga et al., 2023). A similar pattern of alpha-beta power modulation has also been observed during single-word production, where it has been linked to the retrieval of lexical-semantic information (Piai et al., 2020). Yet, whether these similar alpha-beta power dynamics reflect a single vs distinct mechanism is currently debated, and previous work has highlighted potential differences in functional contribution (Zioga et al., 2024) and cortical origin (Cao et al., 2022), despite similar activation profiles. In the context of our experiment, and given the aforementioned theoretical accounts, we speculate that alpha-beta power might conjunctively support the top-down encoding of visual semantic categories and recruitment of their respective networks, which is larger for cues activating more precise visual representations. Yet, future studies are needed to assess the specific roles of alpha and beta oscillations in supporting language-mediated visual object recognition, and elucidate whether these frequencies are acting in conjunction or reflecting different mechanisms.

Despite the present findings remain inconclusive about the specific functional roles of alpha and beta oscillations, as well as their potential dissociation, they provide novel contributions to debate on whether language shapes perception at the early or late stages of perceptual processing. Evidence for the former account comes primarily from EEG studies showing that language affects visual processes by modulating early ERP components such as the P1 (Boutonnet and Lupyan, 2015; Noorman et al., 2018) and N170 (Landau et al., 2010). However, studies focusing on post-stimulus activity are also coherent with a later semantic or decision-making account. Indeed, post-stimulus differences, even if very early, could still emerge from rapid feed-forward integration of visual and linguistic information (Thierry et al., 2009). By showing language-induced modulations of alpha-beta power in posterior regions before image presentation, our findings align with the idea that linguistic influences on visual perception arise at early stages and in a top-down manner.

Finally, our study included a novel manipulation not considered in previous studies on categorization: the inclusion of L2 words as auditory cues. Our participants were highly proficient Basque-Spanish bilinguals, with comparable levels of proficiency in both languages, who had acquired their L2 later in development. The effect of top-down processing in bilinguals is currently debated, and largely dependent on factors like proficiency (Kaan, 2014) and age of acquisition (Molinaro et al., 2017). Although it is commonly believed that bilinguals access a semantic system common to both languages (Caramazza and Brones, 1980), recent studies have suggested that top-down processing may be reduced in a second language because of reduced access to perceptual memory resources (Hayakawa and Keysar, 2018), which are known to play an important role in the generation of visual expectations (Hindy et al., 2016). We found comparable behavioral and neural responses after L1 and L2 words cuing visual object recognition. This result is in line with the idea that both languages provide similar types of top-down guidance to the visual system.

However, our results show that L1 and L2 words both affect visual processing differently than natural sounds, challenging the hypothesis that such cues provide similar top-down semantic information to visual regions. Why do symbolic cues enhance visual object recognition performance more than natural sounds? It has been proposed that symbols are extremely effective in compressing semantic information in a format that transcends within-category differences, thus leading to the amplification of those prototypical features that are relevant for distinguishing between exemplars of different categories (Lupyan and Thompson-Schill, 2012). By contrast, natural sounds are primarily linked to context-specific sources (e.g., the barking of a dog may trigger the representation of a specific exemplar of a dog), thus being less effective at cueing categorical states (Edmiston and Lupyan, 2015). Interestingly, ascribing labels to experiences has also been shown to enhance other cognitive functions, such as the retention of items in visual working memory (Souza and Skóra, 2017), learning novel categories (Lupyan et al., 2007), and perceptual categorization across sensory modalities (Miller et al., 2018). These findings indicate that language acts as a powerful tool for compressing information, facilitating different operations important to a multitude of human cognitive processes (Clark and Toribio, 2012). Future studies should investigate whether similar oscillatory mechanisms are employed to support these language-augmented cognitive functions.